语音分离技术的研究现状与展望

从计算听觉场景分析和盲源分离两种方法综述了当前语音分离技术的研究现状和发展方向。计算听觉场景分析是用计算机来模拟人类听觉系统的处理机制。它可分为两大类：一类是数据驱动型，特点是信息由低级向高级的单向流动；另一类是图式驱动型，特点是信息由低级向高级和由高级向低级结合的双向流动。最后指出信息双向互流的混合语音分离模式将是未来计算听觉场景分析研究的主要模式：基于听觉和视觉的结合来改善语音分离效果的研究将是未来研究方向之一。此外，盲源分离的欠完全问题．非线性混叠信号的可分离性、非平稳混叠信号的盲分离问题都将需要进一步研究：基于CASA和BSS联合进行语音分离将是未来研究的热点。

State and frontiers of research in speech separation

The state and the direction of research in speech separation technique, including computational auditory scene analysis （CASA） and blind source separation （BSS）, are reviewed. Computational auditory scene analysis is a processing mechanism that simulates human auditory system by computer. CASA is classified as data-driven CASA and schema-driven CASA. The information direction flowing of data-driven CASA is bottom-up and that of schema-driven CASA is integrating bottom-up and top-down. Finally, it is pointed out that the speech separation model of information bi-direction flowing of CASA is the dominant pattern of CASA in the future, and the speech separation method based on integrating audition and vision in order to improve effect of speech separation is one of the directions of research in the future. Furthermore, it is pointed out that the under-complete, the nonlinear and the non-stationary BSS problems are needed to study further, and the speech separation technique based on integrating CASA and BSS is the hot issue of research in the future.